26/07/2022, AI
If you follow me on Twitter, you’re probably aware I have become slightly obsessed with OpenAI’s GPT-3, a machine learning language model. I have released several free Google Sheets scripts for marketers utilising this technology.
I also recently gained access to OpenAI’s Dalle-2, a new AI system that creates realistic images and art from natural language descriptions. It offers a feature with which you can upload an existing image and ask Dalle-2 to add to it, change it or generate variations.
There has been some incredible artistic work produced with this tool. However, I am more interested in its practical applications as a business tool.
I have included below a few examples below and will add to these as new ideas spring to mind.
Experiment 1) – Swimming Pools
Imagine you own a swimming pool installation business. You want to help your prospective clients visualise their garden with the pool installed.
1) Take a picture of their garden and upload it to Dalle-2.
2) Add a rectangular pool with a stone surround.
3) Add a raised patio with some garden furniture.
4) Add a family swimming in the pool.
5) Add a small pump room and make it a sunny day.
Take a picture of the rear of a house and ask Dalle-2 to generate multiple variations.
Original image
Variation 1
Variation 2
Variation 3
Upload an image of a house and ask Dalle-2 to add a loft conversion with Dorma windows.
Original image
With loft extension added
Upload an image of a garment. Ask Dalle-2 to display it on a model (only partially successful).
Original image
Displayed on a model (needs more work)
Upload an image of a flower bed. Ask Dalle-2 to generate ideas around the existing flowers.
Original image
Variation
I’m not qualified to give a technical answer to this question, and you probably don’t want one.
What I can tell you is that the images generated by Dalle-2 are not just “stitched together from the web”. It has, however, been trained on billions of existing digital images.
Imagine I asked you to draw a picture of “a dog with the neck of a giraffe”. You would probably recall in your mind an image of a dog and a giraffe and just keep drawing strokes on paper until you produced something resembling the instruction to the best of your artistic ability.
Dalle-2 works in a similar way. It “understands” the properties of a dog and a giraffe. It then generates variations until satisfied that the output shares enough of those same properties.
There is currently a waiting list to use Dalle-2 but applications are being processed quickly. New users receive 50 image credits free and 15 more each month. You can also buy credits in units of 115 for $15 (460 images).
Dalle-2 is impressive, but far from perfect. You will waste a lot of credits on failed generations. However, you will get better results once you start to understand its quirks and limitations.
I feel I have barely scratched the surface of the capabilities of this technology. I would love to hear your ideas or suggestions.
Check out some of my other scripts and tools for marketers using OpenAI’s technology.